apple-on-device-ai
Pass
Audited by Gen Agent Trust Hub on Aug 23, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill provides educational content and implementation examples for Apple's Foundation Models, Core ML, MLX Swift, and llama.cpp. All code examples are standard Swift and Python for these frameworks.
- [SAFE]: The skill contains a dedicated section on safety and guardrails, explicitly instructing developers to treat model output as untrusted and to avoid executing side-effects directly from generated text.
- [EXTERNAL_DOWNLOADS]: The documentation references official repositories like github.com/ml-explore/mlx-swift-examples and mentions downloading model weights from Hugging Face. These are well-known and reputable sources in the AI/ML ecosystem.
- [PROMPT_INJECTION]: The skill addresses indirect prompt injection risk by providing a clear evidence chain: (1) Ingestion points include user-provided prompt text and tool arguments in the respond(to:) and call(arguments:) methods. (2) Boundary markers are implemented via the instructions parameter in LanguageModelSession. (3) Capability inventory includes tool execution, local model loading, and external weight downloads. (4) Sanitization is explicitly recommended, advising users to treat all generated text as untrusted and to validate actions independently.
Audit Metadata